fix: use tl.split instead of indexing for E2M1 pair packing

Triton doesn't support constexpr tensor indexing (e2m1_pairs[:, 0]).
Use tl.split() which splits the last axis into two tensors.

patches/deepseek_v4.py

@@ -353,9 +353,8 @@ def _deepseek_v4_stage_mega_moe_inputs_kernel(
     # Pack 2 E2M1 values per byte: even→low nibble, odd→high nibble
     PACKED_K: tl.constexpr = BLOCK_K // 2  # 64
     e2m1_pairs = tl.reshape(e2m1_4bit, [PACKED_K, 2])
-    even = e2m1_pairs[:, 0].to(tl.uint8)
-    odd = e2m1_pairs[:, 1].to(tl.uint8)
-    packed_byte = (odd << 4) | even
+    even, odd = tl.split(e2m1_pairs)  # splits last axis (size 2) into two [PACKED_K] tensors
+    packed_byte = (odd.to(tl.uint8) << 4) | even.to(tl.uint8)
 
     packed_k_offsets = k_block_id * PACKED_K + tl.arange(0, PACKED_K)
     packed_k_mask = packed_k_offsets < (hidden_size // 2)